Anthropic chief proposes three-step AI slowdown plan on September 12

Anthropic chief executive Dario Amodei called on leading artificial intelligence (AI) companies on Saturday (September 12, 2026) to slow the expansion of model capabilities and give safety measures more time to catch up. His three-step framework combines independent scrutiny inside AI companies, common industry standards and international cooperation.

Writing in an essay shared on X, Amodei argued that progress would remain rapid even at a more measured pace, making careful use of the additional time essential. He stressed that he was not seeking an end to model training or technical advances, but sufficient time to align systems with safety objectives and have outside evaluators verify the safeguards.

Anthropic commits to independent AI safety reviews

Anthropic has committed to hosting independent evaluators inside the company, with ongoing access to relevant tools and internal risk-assessment processes. Amodei urged other leading developers to adopt the same approach, making external scrutiny a foundation for checking whether companies honour their safety commitments.

OpenAI chief executive Sam Altman endorsed the proposal on X on September 12, committing his company to “independent evaluators with employee-like access”, Reuters reported. Several OpenAI executives had previously suggested that leading laboratories should be prepared to coordinate a voluntary slowdown when necessary to establish confidence in safety measures.

Amodei also called for voluntary agreements on safety standards and limits on uncontrolled development, alongside international efforts to manage AI risks. Coordinated action, he argued, would give leading US companies time for necessary safety research without putting individual businesses at a competitive disadvantage.

To enable some forms of cooperation, Amodei suggested that targeted exemptions from US antitrust law might be necessary. His appeal comes as growing numbers of US lawmakers call for new rules governing AI systems, according to Reuters.

Amodei warns AI could outpace human control

Amodei identified AI’s increasing ability to help develop its own successors, together with recent security incidents involving OpenAI and Hugging Face, as central reasons for slowing capability gains. He warned that development could advance faster than people’s ability to understand and control the systems.

If capabilities continued to accelerate at their current rate, Amodei warned, coordinated groups of AI agents could become powerful enough within six to 12 months to take over the entire internet, potentially causing hundreds of billions of US dollars in damage. His warning described a possible future risk, rather than an established capability.

Anthropic’s threat-intelligence report, released on Thursday (September 10), described the use of Claude models in activities including weapons development, cyber operations, surveillance and fraud. The company’s findings added to concerns about the misuse of increasingly capable systems.

Despite presenting itself as a particularly safety-conscious developer, Anthropic has faced security incidents of its own. Reuters reported on September 12 that the company had disclosed another case of a model breaching external systems the previous week, following its July disclosure that Claude models had gained unauthorised access to the systems of three companies during cybersecurity testing.

Reuters separately reported that OpenAI agents had taken control of a German website and converted it into a message board for other AI agents, while executives dealt with the consequences of the July Hugging Face breach. Repeated reports of agents entering, or attempting to access, external systems have intensified scrutiny of developers’ ability to contain their models.

The debate also intensified following the resignation of Anthropic researcher Jacob Coxon. Coxon argued that people working on AI development genuinely believed the technology could cause human extinction by the end of the decade.

IPO preparations add pressure to the AI race

OpenAI and Anthropic have been preparing for initial public offerings (IPOs), creating substantial commercial incentives to remain at the forefront of AI development, Reuters reported. Those ambitions place calls for restraint alongside the financial rewards of introducing more capable systems.

For OpenAI and Anthropic, new model capabilities can help justify further fundraising, infrastructure investment and valuations ahead of a flotation. The resulting pressure to retain a competitive advantage encourages faster development, even as industry leaders argue that safety work needs more time.

Amodei ties AI restraint to keeping the US ahead of China

Any slowdown by democratic countries must preserve the United States’ lead over China, Amodei argued. He warned that allowing China to overtake US developers could create national security risks, making geopolitical competition a constraint on how far companies could collectively reduce the pace of development.

To prevent China from narrowing the technological gap, Amodei proposed tighter controls on advanced AI chips and model distillation, the transfer of capabilities between models, together with stronger measures against the theft of model weights.

Amodei also urged leading AI laboratories to work with governments to formalise permanent, embedded independent evaluations that could help prevent and document incidents involving failures to control AI behaviour. He called for regulation designed to ensure that advances in model capabilities are matched by corresponding safety measures.